A CONCEPT · PROJECTED FROM THE VINDEX KNOWLEDGE GRAPH
KV STATE vs RECURRENT STATE
growing memory or folded memory
Two kinds of continuation memory: softmax attention keeps every past key and value and grows with context; linear families fold the past into a fixed-size state.
A hybrid stack carries both at once, layer by layer, and the plan declares which — row geometry and windows for KV layers, state shape for recurrent layers. MLA sits between them: true KV state, stored through a latent bottleneck. Mamba2 sits wholly on the folded side: SSM state plus a short conv history, and no KV row anywhere in the model. The runtime's rule covers all of it: state geometry is a container fact.
IN THE GRAPH
kv state vs recurrent state— distinguishes →recurrent state
kv state vs recurrent state— belongs to →attention